Knowledge Graph and Large Language Model Fusion for Explainable Medical Question Answering

Abstract

Medical question answering has moved from retrieval of curated documents toward generative systems that synthesize evidence across heterogeneous clinical resources. Large language models offer fluent, context-sensitive responses, yet their probabilistic nature creates risks of hallucination, citation drift, and opaque reasoning. Knowledge graphs provide structured, auditable representations of biomedical entities and relations, but they are often incomplete, brittle, and disconnected from natural language interaction. This paper examines the fusion of knowledge graphs and large language models for explainable medical question answering as a socio-technical systems problem rather than a purely model-centric task. It analyzes architectural patterns that couple graph retrieval, semantic alignment, attention-based fusion, and constrained generation, with attention to trade-offs among expressiveness, latency, verifiability, and maintainability. The discussion then turns to explainability as an infrastructure requirement, arguing that clinical explanations must connect model outputs to provenance, evidence hierarchies, and workflow-specific decision thresholds. Governance, privacy, regulatory, and sustainability concerns are examined through the lens of deployment across hospitals, research networks, and patient-facing services. The paper also assesses fairness and robustness risks arising from incomplete graphs, biased documentation, and unequal data representation. It concludes that explainable medical question answering requires layered validation, human oversight, transparent data lineage, and policy mechanisms that treat knowledge graphs and language models as continuously governed clinical infrastructure. The central contribution is a system-level agenda for designing fusion architectures that are clinically useful, institutionally accountable, and resilient under real-world uncertainty.